Lifecycle Copy — Channel Copywriting

Produce send-ready copy for journey steps. Copy here is an engineering artifact: it has hard constraints (limits, banned words, consent text) and is reviewed before it is shown. The mandatory format is ${CLAUDE_PLUGIN_ROOT}/templates/copy-output.md.

When NOT to use this

• No journey doc or step description exists yet — there's nothing to write copy for; run lifecycle-journeys first or supply a step description (channel + intent) directly.
• The ask is whether EXISTING, already-shipped copy complies with channel/lexicon rules — that's lifecycle-audit's dimension 7 (spot-check), not this skill; this skill's review loop only covers copy it itself writes.
• The need is a CRM-agnostic structured export of copy already written — that's lifecycle-export, not this skill.

Inputs (gate)

1. A journey doc (from lifecycle-journeys) OR a user-supplied step description (channel + intent minimum).
2. The rule chain, loaded in inheritance order (CLAUDE.md rule 8 — most specific wins, bans/compliance strictest wins):
   • ${CLAUDE_PLUGIN_ROOT}/knowledge/brands/<brand>.md if it exists — tone/formality override; extra_banned_words extend all ban lists; brand_vocabulary is authoritative for product/feature names.
   • ${CLAUDE_PLUGIN_ROOT}/knowledge/lexicons/locales/<lang>.md per target language — voice, emotion calibration, market red lines.
   • ${CLAUDE_PLUGIN_ROOT}/knowledge/lexicons/<sector>.md. No lexicon for the sector → use the closest and say so.
3. Tone + formality + language(s) from the brand file or Intake Summary. Missing → trigger lifecycle-intake (tone questions only).
4. Persona context: the journey doc's §3 audience definition (segment, lifecycle stage, RFM tier where known) is passed to the writer — copy addresses that persona, never a generic user.
5. Failed-strategies check: if output/<brand>/failed-strategies.md exists, a logged powered failure is not re-proposed to the same segment; when a strategy is skipped because of the log, say which entry caused it.
6. Winning-strategies check: if output/<brand>/winning-strategies.md exists and has a confirmed entry for this segment/journey stage, pass it to the writer as a precedent — it informs one variant's starting angle, never both; the other variant still explores a genuinely different, less-tested angle. A precedent narrows a starting point, it doesn't replace the two-distinct-angles rule.

Procedure — write → review → fix (mandatory loop)

1. Load the rules

For every channel used, load ${CLAUDE_PLUGIN_ROOT}/knowledge/channels/<channel>.md — its frontmatter limits and "Hard rules" are the review contract. Load the lexicon's use/avoid table, urgency rules, and banned list.

2. Write

When multiple journeys need copy, spawn one copy-writer per journey concurrently in a single message (the agent is designed for exactly this parallelization). Pre-bundle the shared rule files (the sector lexicon, each channel file's limits + hard rules, the locale overlay, and the brand config's tone/banned/vocabulary) into each writer's prompt verbatim, read once here rather than re-read by every writer — the same cost saving the journey stage applies, and it matters more here because copy fans out to two agents per journey (writer + reviewer). Each writer then only needs its own journey doc, not a fresh read of the identical rulebook. For each journey step (its intent line is the brief — don't drift from it):

• A target tone per step, drawn from the lexicon's stage-calibration table and adjusted by the locale overlay's emotion calibration — goes in the template's Target tone field and binds the writer.
• Variant A and Variant B with genuinely different angles (e.g. utility vs social proof) — not synonym swaps. Each variant carries the template's strategy + hypothesis JSON metadata (one falsifiable sentence — lifecycle-results scores these later).
• One short Fallback with no personalization variables.
• Real character counts per field (count characters, do not estimate; for SMS respect the GSM-7 vs Unicode distinction in the channel file — Turkish characters change the limit).

3. Review (adversarial)

Delegate to the copy-reviewer agent (or, if subagents are unavailable, perform the same checklist yourself in a separate explicit pass):

• limits per field · banned/spam words (channel + lexicon) · single-CTA rule · urgency claims backed by real data variables · variable fallbacks exist · consent/opt-out text where required · tone matches lexicon calibration for the journey stage.
• Verdict per block: PASS / FIX (with the exact violation).

4. Fix and re-check

FIX blocks are rewritten and re-checked. Nothing labeled FIX may reach the user. If a rule cannot be satisfied (e.g. intent impossible in 120 push chars), change the step's channel recommendation and tell the user why.

Legal-review path: when the sector lexicon has regulated: true and a block carries a borderline money/health/outcome claim that is not outright banned, the block ships stamped ⚖️ legal review required (with the exact claim quoted) instead of ✅ — it goes to manual sign-off, not into the send queue.

5. Deliver

Mandatory format: ${CLAUDE_PLUGIN_ROOT}/templates/copy-canvas.html, reproduced verbatim via python3 scripts/build_canvas.py --template ${CLAUDE_PLUGIN_ROOT}/templates/copy-canvas.html --journeys <journeys.json> --meta <meta.json> --out output/<project>/<name>.html — the same deterministic copy-and-swap the journey canvas uses (it substitutes only the JOURNEYS data array, <title>, and eyebrow/h1/lede, self-verifies no boilerplate drifted, and handles copy-canvas's absence of HOLDOUT_TIP/DATA_NOTE). Hand-editing the template is only a fallback if the script is unavailable; never retype the boilerplate. The artifact's user-facing name and header follow the user's language, and never use the word "copy" toward Turkish users (it reads as "kopya/duplicate"): TR → "İletişim Metinleri", EN → "Message Copy". This covers the generated FILE NAME too: write the output to output/<project>/iletisim-metinleri.html (TR) / message-copy.html (EN) — and the markdown source-of-truth alongside it with the same base name — never copy-canvas.html/copy-*.md; only the repo template keeps its English name. One journey per tab, one card per step, review-status badge per card (✅ / ⚠️ with note / ⚖️ legal). The plain templates/copy-output.md structure is still the underlying content model each step's fields/variants/fallback follow — the HTML canvas is how it reaches the user, per this repo's output-authoring rule (CLAUDE.md rule 2). Multi-language requests: write each language natively per the lexicon's TR/EN notes — never translate literally.

Word choice discipline

• The lexicon decides vocabulary; if a lexicon rule and a "nicer sounding" line conflict, the lexicon wins.
• Concrete beats clever: product names, real numbers (as {{variables}}), stated policies.
• Urgency/scarcity only when backed by a data variable that exists ({{stock_count}}, {{sale_end_date}}). No data → no urgency. Ever.

Common Pitfalls

Pitfall 1: Variants that differ in wording, not in angle. Symptom: Variant A and B say the same thing with synonyms swapped — "Sepetin seni bekliyor" vs "Sepetindeki ürünler seni bekliyor." Consequence: the two variants test the same hypothesis twice; lifecycle-results has nothing to attribute a lift to, because there was only ever one lever in the test. Fix: check each variant's strategy field before delivery — if both name the same lever (both "urgency," both "utility"), rewrite one around a genuinely different lever (social proof, loss-aversion, utility) rather than rephrasing.

Pitfall 2: Urgency framing with no variable behind it. Symptom: a line reads urgent ("Fırsat kaçmadan şimdi bak") but no {{stock_count}}, {{sale_end_date}}, or other real data variable appears anywhere in the field. Consequence: this is the "no data → no urgency" rule violated by tone rather than by an obviously banned word — a literal banned-word check won't catch phrasing that implies scarcity without stating it. Fix: trace any urgent-reading line to the specific variable driving it; no variable found means rewrite without the pressure framing, not just soften the punctuation.

Pitfall 3: A FIX rewritten but not re-checked. Symptom: a block flagged FIX gets rewritten to fix the named violation, then shipped on the strength of "it reads better now." Consequence: the rewrite can introduce a new violation the first pass never had — shortening for a length limit can drop the required opt-out text, tightening a CTA can create a second one. Fix: every FIX goes through the full checklist again after rewriting, not a visual read — the loop is write → review → fix → review, not write → review → fix → ship.

Never do

• Never show unreviewed copy or skip the review pass "because it's short".
• Never output copy without character counts, or with estimated counts.
• Never use a personalization variable that the user's data cannot fill (check the event/attribute inventory).
• Never write identical copy across variants or across channels (each channel has its own shape, not a resize).
• Never use an em dash or en dash in customer-facing copy — commas, periods, parentheses, conjunctions.
• Never ASCII-fold diacritics (ç, ğ, ı, İ, ö, ş, ü …) in any channel; the only exception is deliberate, user-approved GSM-7 folding for SMS cost (see knowledge/channels/sms.md).